Two lines that still add up
Suppose a buyer sends a manufacturer a request for quotation (RFQ) for two versions of a steel bracket. B-17L is the left-hand bracket and B-17R is its right-hand mate. An AI intake tool reads the attachment and prepares this draft for estimating.
| Part | RFQ | Draft |
|---|---|---|
| B-17L | 24 each | 240 each |
| B-17R | 240 each | 24 each |
Both part numbers exist. Every digit came from the buyer. The total is 264 brackets on either side. An estimator who sees only the draft could price a long run for the left-hand part and a short run for the right-hand one. If setup time matters, the swap can change both unit prices.
This is a useful first test for manufacturing teams connecting intake with estimating. A document can yield the right words and numbers while losing the relationship a buyer intended. The acceptance test needs to catch that loss before anyone prices the work.
Make the quote line the unit of review
Write down what the estimator needs to price one line. For the brackets, that means the buyer’s line identifier if present, the part number including its suffix, the requested quantity, and the purchasing unit. If a drawing revision changes what the shop would make, include the revision. Keep these values together from the moment the system proposes a line through the handoff to estimating.
Keep a way back to the exact attachment version and location that support the proposal. A page reference helps. An outline around the part, quantity, and relevant column header helps more, especially when adjacent rows differ by one character. For example, Amazon Textract returns row and column indexes and the position of a detected cell on the page. Those coordinates give the reviewer somewhere to look. The estimator still decides whether the cells belong together. Amazon Textract table documentation.
Let a blank unit stay blank
Suppose the next RFQ lists “4 boxes” for one part and just “4” for the next. First check whether a shared column heading supplies the unit for both. If the second line still has no unit, keep its quantity, leave the unit blank, and ask the buyer what they meant. Even when “boxes” is explicit, converting boxes to pieces requires a verified pack size for that part. Keep the buyer’s original wording alongside any conversion so an estimator can retrace it.
Give the estimator the evidence at the decision
Imagine opening the draft and selecting B-17L. Show the buyer’s row with the neighboring row and column heading still visible. Cropping down to a picture of “240” would make the incorrect quantity look convincing. Let the estimator correct the proposed line, see which other line might be affected, and check both before releasing the quote for pricing.
Some exceptions have a narrow scope. A missing unit on line seven can leave the first six lines available for preparation. Others change the whole request. If a note says “quote as one lot,” hold the package until the open line is resolved. Agree on that rule with the person who releases the quote. An operator should be able to tell which work can continue and what would send an incomplete price to a buyer. Vilcorp’s guide to review states for AI intake walks through assigning those decisions to a person.
Score the pairs an estimator would accept
Make a small answer key with an estimator before the pilot. Give each expected line its correct part, quantity, unit, and any required revision. Count a match only when the draft assigns all of them to the same line. Keep missing lines, unexpected extra lines, and requests sent for clarification visible as separate outcomes. That way a tool cannot earn a good score simply by offering fewer lines for review.
Start with three variations of the bracket request.
- Swap the two quantities in a candidate draft. The check should flag both rows even though every number appears in the RFQ and the total balances.
- Move the buyer’s second row to the next page. Repeat the column header. Expect two quote lines, each linked to the correct page, with no extra line for the header.
- Ask for two price breaks on one part. Request separate prices for 24 and 240 pieces of B-17L. The draft should preserve the two alternatives. Adding them into a single order for 264 pieces changes the question the buyer asked.
Use the same examples when the extraction tool, model, or document template changes. Vilcorp’s guide to building an AI evaluation set explains how to keep an agreed set of cases in the release process. Alongside the scores, time how long an estimator needs to reach a usable draft, including the time spent opening attachments and fixing errors. Those minutes tell you whether automation has actually removed work.
Decide what can enter pricing
For a first release, pick one family of incoming RFQs and run the drafts beside the existing intake process. Have the estimating lead inspect swapped, missing, and ambiguous lines and decide which layouts the team can support. Keep an unfamiliar layout in manual intake until someone has tested it. A buyer waiting for a price can work with a clear question about a unit. A polished quote for the wrong bracket is harder to unwind.
When planning AI automation for an existing workflow, bring one representative request and the draft an estimator would accept. Put them side by side and ask someone to point to the source of each priced line. If the numbers add up but the parts have traded places, you have found the first test to put in the release.
